Eight Weeks of AI Training Rates: What Actually Moved (And What We Got Wrong)

2026-08-27 · Expert Match AI team

Most pay figures in AI training work are snapshots. Somebody surveys a forum, or quotes a platform's marketing page, and the number sits there uncontested for a year.

We have been recomputing ours every six hours since July 2, 2026. That is now eight weeks of continuous rate history on the same pipeline, which is short enough to be honest about but long enough to show movement - and, this edition, long enough to catch us being wrong about something.

Start with the thing we got wrong

On August 11 we split the board by when each listing first appeared and reported this:

| Cohort (as of Aug 11) | Priced roles | Avg rate | |---|---|---| | First seen before August | 134 | $87/hr | | First seen in August | 92 | $106/hr |

We called that "the single most useful number in this post" and said the market was being restocked at a higher price point.

Sixteen days later, the same calculation on the same pipeline says the opposite:

| Cohort (as of Aug 27) | Priced roles | Avg rate | |---|---|---| | First seen before August | 58 | $112/hr | | First seen in August | 151 | $84/hr |

The August cohort is now arriving 25% below the older inventory, not 22% above it.

Two things happened, and only one of them is about pay. The first is survivorship: the pre-August cohort has shrunk from 134 listings to 58, and the ones that survived are disproportionately the premium evergreen tiers - Surge, Mercor talent networks, Braintrust senior roles - that stay open for months. Cheap listings churn faster, so an ageing cohort *looks* like it is getting richer while it is really just getting smaller. The second is that August's restock, taken across the full month rather than its first eleven days, was mostly mid-band micro1 volume.

The lesson we would draw is the one we keep drawing: an average with a shrinking denominator is not a price signal. We printed a 22% premium as a forward-looking indicator two weeks ago. It was mostly an artefact of a cohort that had not finished churning. That is the kind of thing eight weeks of data catches and eleven days of data does not.

Field-by-field, across all three snapshots

| Field | Jul 31 | Aug 11 | Aug 27 | Net move | |---|---|---|---|---| | Legal | $138 | $218 | $196 | +$58 | | Science | $70 | $104 | $111 | +$41 | | Finance | $72 | $77 | $98 | +$26 | | Business | $43 | $93 | $70 | +$27 | | Linguistics | $46 | $55 | $58 | +$12 | | Software | $84 | $87 | $92 | +$8 | | ML, Data & AI | $65 | $70 | $63 | −$2 | | Healthcare | $136 | $116 | $119 | −$17 |

Before reading anything into that, note the mechanic that drives nearly all of it: on a board this size, averages move because of composition, not because anyone renegotiated a rate. No platform cut anybody's pay this summer. What changed is which listings exist.

Legal: the spike was real, and it has partly retraced

Legal's climb from $138 to $218/hr in early August was the largest single move we have recorded, and it traced to one source: micro1 opening a block of attorney specialisations at $140–$400/hr each.

Those roles are still live - BigLaw, corporate, family law, immigration, litigation associate and M&A - and micro1 has since added a further tier of litigation attorney roles at $100–150.

That second wave is why the field average fell back $22. Not because the $400 band went away, but because more legal work arrived *underneath* it. Legal now carries 22 priced roles against 15 in July, and the field has roughly doubled in depth while giving up a tenth of its headline rate. If you are an attorney, more doors at $100–150 is a better market than fewer doors at $400, whatever the average says.

So: was the spike a real market signal or a data artefact? Eight weeks in, the answer is that it was a real signal about *demand* and a bad guide to *rates*. micro1 does not open a dozen simultaneous legal queues speculatively. But the number attached to that decision was never going to hold at $218.

Healthcare: the decline reversed itself, quietly

Healthcare fell $20/hr in early August and has since ticked back up $3, to $119.

The August dip had a clean explanation at the time: a wave of mid-tier clinical listings - medical coders, medical auditors, regulatory affairs specialists, clinical document authors, all in the $50–80 band - joined a field whose average had rested on a handful of premium clinician tiers. The average fell because the field got deeper, which was the *good* outcome.

Most of that wave is no longer live. The premium anchors are unchanged and still open: Surge's physician tier at $250–450 and Mercor's Physician Talent Network at $110–250. Underneath them sit Braintrust's RLHF Healthcare Expert at $100–180 and micro1's hospitalist roles at $70–100.

This is worth remembering every time you see a pay index quoted without its composition: healthcare has now been down $20 and up $3 in seven weeks without a single rate changing.

Science and finance: the quiet winners

Science climbed $41/hr across the eight weeks to $111, and now sits third on the board. Unlike legal, this was not one employer - it is a broad restock of physics, chemistry, psychology and research-expert roles across micro1, Mercor and Turing, topped this week by a PhD & Academic Expert listing at $245–280.

Finance is the surprise of the month: +$26 to $98, almost all of it from a single new Investment & Finance Expert role at $245–280/hr. Ten priced roles is a small enough field that one listing moves it twenty dollars, so treat that with the scepticism it deserves - but it is the first time finance has broken out of the $70s since we started tracking.

By platform

| Platform | Jul 31 | Aug 11 | Aug 27 | |---|---|---|---| | Surge | $406 | $406 | $406 | | Mercor | $102 | $118 | $136 | | Braintrust | $105 | $105 | $105 | | Stellar | $87 | $87 | $87 | | micro1 | $65 | $90 | $83 | | Alignerr | $64 | $64 | $64 | | Mindrift | $71 | $69 | $58 | | DataAnnotation | $52 | $52 | $52 | | Outlier | $32 | $32 | $32 |

Five of nine platforms have not moved once in eight weeks. Surge, Braintrust, Stellar, Alignerr, DataAnnotation and Outlier are posting the same bands they posted on July 2. The entire board-level story is micro1 and Mercor - and micro1 carries 144 of the priced listings, so when it moves, the market moves.

The 12x platform spread we flagged in July is essentially unchanged: Surge to Outlier is still about 12.7x. Strip Surge out as the outlier it is, and the rest of the board spans a narrower 4.3x from Mercor to Outlier. Where you apply continues to matter more than what you know - see why serious trainers work 2–3 platforms at once.

What we would tell you to watch

Whether micro1's legal band holds. A dozen roles at $140–400, now with a $100–150 tier underneath them, is either a market finding its level or an opening bid settling lower. Eight weeks still is not long enough to know, but the direction is now visible.

Mercor's climb. Up $34/hr across eight weeks, entirely on talent-network listings with wide bands. It is the only platform besides micro1 doing anything at all.

Composition, always. Every significant move in this post was caused by listings arriving or expiring, not by rates changing. Any pay figure in this market - including ours, especially ours - should be read with the role count next to it. We printed a cohort premium in August that inverted within a fortnight; that is what this failure mode looks like in practice.

Methodology

Averages use the midpoint of each listing's disclosed hourly range, across every priced role live at snapshot time; per-project, annual and undisclosed-rate roles are tracked but excluded from every figure here rather than converted. Cohort figures use each listing's first-seen timestamp in our pipeline, which begins 2026-07-02 - roles that existed on platforms before that date are dated from when we first observed them, so "first seen in August" means new to us, not necessarily new to the world.

July and August 11 comparison figures come from the snapshots published in the July monthly report and our August 11 analysis. August 27 figures were computed against the live board on the afternoon of August 27; the salary report shows slightly different values because it recomputes on a six-hour cycle and its most recent pull predates several listings counted here. Small fields swing on single listings; healthcare, finance, science and business all have fewer than 16 priced roles, and Mindrift's figure comes from the salary report because its listings quote a ceiling with no floor.

Live numbers recompute every six hours on the salary report, and every role behind this analysis is on the jobs board.

Published by the Expert Match AI team. Pay figures are listed rates from live postings, not a survey of accepted offers. Some outbound application links on this site carry disclosed referral codes; rankings and recommendations are never influenced by them.